/* * OPTIMIZED INT16 DDC IMPLEMENTATION - 33-TAP VARIANT * * Based on commit https://github.com/airspy/airspyone_host/commit/091c6f7449bc976f97bb327242532bfb1001d116 * * Original work Copyright (C) 2014, Youssef Touil * Modifications Copyright (C) 2025-2026, Benjamin Vernoux * * SPDX-License-Identifier: MIT */ #include "iqconverter_int16_opt33.h" #include "iqconverter_common.h" #include "iqconverter_lut.h" #include #include /* 33-tap filter: 8 symmetric pairs, power-of-2 delay for branchless modulo */ #define SYM_TAPS 8 #define DELAY_SIZE 8 #define DELAY_MASK 7 iqconverter_int16_opt33_t *iqconverter_int16_opt33_create(const int16_t *hb_kernel, int len) { iqconverter_int16_opt33_t *cnv; int i; if (len != 33) return NULL; cnv = (iqconverter_int16_opt33_t *)iqconv_aligned_alloc(sizeof(iqconverter_int16_opt33_t)); if (!cnv) return NULL; memset(cnv, 0, sizeof(iqconverter_int16_opt33_t)); /* * Extract symmetric kernel coefficients (every other tap, first half) * For 33-tap filter, non-zero coefficients are at ODD indices (1,3,5...) */ for (i = 0; i < SYM_TAPS; i++) cnv->sym_kernel[i] = (int32_t)hb_kernel[i * 2 + 1]; cnv->dly_len = DELAY_SIZE; return cnv; } void iqconverter_int16_opt33_free(iqconverter_int16_opt33_t *cnv) { if (cnv) iqconv_aligned_free(cnv); } void iqconverter_int16_opt33_reset(iqconverter_int16_opt33_t *cnv) { if (OPT_UNLIKELY(!cnv)) return; cnv->fir_head = 0; cnv->dly_head = 0; cnv->old_x = 0; cnv->old_y = 0; cnv->old_e = 0; memset(cnv->fir_buf, 0, sizeof(cnv->fir_buf)); memset(cnv->dly_buf, 0, sizeof(cnv->dly_buf)); } /* * FUSED PROCESSING: uint16_t ADC samples -> int16 I/Q output * Uses precomputed LUT for ultra-fast ADC -> int16 conversion. */ OPT_HOT void iqconverter_int16_opt33_process_u16(iqconverter_int16_opt33_t * OPT_RESTRICT cnv, const uint16_t * OPT_RESTRICT src, int16_t * OPT_RESTRICT dest, int len) { int i; int16_t ox = cnv->old_x; int16_t oy = cnv->old_y; int32_t oe = cnv->old_e; int32_t f_head = cnv->fir_head; int32_t d_head = cnv->dly_head; /* Apply alignment hints to enable better vectorization */ int32_t * OPT_RESTRICT fir_buf = (int32_t *)OPT_ALIGNED_64(cnv->fir_buf); const int32_t * OPT_RESTRICT kernel = (const int32_t *)OPT_ALIGNED_64(cnv->sym_kernel); int16_t * OPT_RESTRICT dly_buf = (int16_t *)OPT_ALIGNED_64(cnv->dly_buf); const int16_t * OPT_RESTRICT lut = (const int16_t *)OPT_ALIGNED_64(iqconv_lut.lut_int16); /* Prefetch first batch of LUT entries */ OPT_PREFETCH_READ(&lut[src[0]]); OPT_PREFETCH_READ(&lut[src[1]]); for (i = 0; i < len; i += 4) { /* Prefetch next batch of LUT entries for next iteration */ if (OPT_LIKELY(i + 8 < len)) { OPT_PREFETCH_READ(&lut[src[i + 4]]); OPT_PREFETCH_READ(&lut[src[i + 5]]); OPT_PREFETCH_READ(&lut[src[i + 6]]); OPT_PREFETCH_READ(&lut[src[i + 7]]); } int16_t x0, x1, x2, x3; int16_t y0, y1, y2, y3; int32_t *fir_src; int32_t acc0, acc1; int16_t q_in0, q_in1; int16_t q_out0, q_out1; int16_t w, s; int32_t u; /* FUSED: LUT-based uint16_t -> int16 + DC Removal */ x0 = lut[src[i + 0]]; w = x0 - ox; u = oe + (int32_t)oy * DC_ALPHA_Q15; s = (int16_t)(u >> 15); y0 = w + s; oe = u - ((int32_t)s << 15); ox = x0; oy = y0; x1 = lut[src[i + 1]]; w = x1 - ox; u = oe + (int32_t)oy * DC_ALPHA_Q15; s = (int16_t)(u >> 15); y1 = w + s; oe = u - ((int32_t)s << 15); ox = x1; oy = y1; x2 = lut[src[i + 2]]; w = x2 - ox; u = oe + (int32_t)oy * DC_ALPHA_Q15; s = (int16_t)(u >> 15); y2 = w + s; oe = u - ((int32_t)s << 15); ox = x2; oy = y2; x3 = lut[src[i + 3]]; w = x3 - ox; u = oe + (int32_t)oy * DC_ALPHA_Q15; s = (int16_t)(u >> 15); y3 = w + s; oe = u - ((int32_t)s << 15); ox = x3; oy = y3; /* FIR I-Channel (Filter 0) - 16-tap symmetric (8 pairs) */ f_head = (f_head - 1) & FIR_BUF_MASK; fir_buf[f_head] = (int32_t)((int16_t)(-y0)); fir_buf[f_head + FIR_BUF_SIZE] = fir_buf[f_head]; fir_src = &fir_buf[f_head]; acc0 = kernel[0] * (fir_src[0] + fir_src[15]) + kernel[1] * (fir_src[1] + fir_src[14]) + kernel[2] * (fir_src[2] + fir_src[13]) + kernel[3] * (fir_src[3] + fir_src[12]) + kernel[4] * (fir_src[4] + fir_src[11]) + kernel[5] * (fir_src[5] + fir_src[10]) + kernel[6] * (fir_src[6] + fir_src[9]) + kernel[7] * (fir_src[7] + fir_src[8]); /* FIR I-Channel (Filter 1) - 16-tap symmetric (8 pairs) */ f_head = (f_head - 1) & FIR_BUF_MASK; fir_buf[f_head] = (int32_t)y2; fir_buf[f_head + FIR_BUF_SIZE] = fir_buf[f_head]; fir_src = &fir_buf[f_head]; acc1 = kernel[0] * (fir_src[0] + fir_src[15]) + kernel[1] * (fir_src[1] + fir_src[14]) + kernel[2] * (fir_src[2] + fir_src[13]) + kernel[3] * (fir_src[3] + fir_src[12]) + kernel[4] * (fir_src[4] + fir_src[11]) + kernel[5] * (fir_src[5] + fir_src[10]) + kernel[6] * (fir_src[6] + fir_src[9]) + kernel[7] * (fir_src[7] + fir_src[8]); /* Mix & Delay Q-Channel - BRANCHLESS with power-of-2 mask */ q_in0 = (int16_t)((-y1) >> 1); q_in1 = (int16_t)(y3 >> 1); q_out0 = dly_buf[d_head]; dly_buf[d_head] = q_in0; d_head = (d_head + 1) & DELAY_MASK; q_out1 = dly_buf[d_head]; dly_buf[d_head] = q_in1; d_head = (d_head + 1) & DELAY_MASK; /* Output */ dest[i + 0] = (int16_t)(acc0 >> 15); dest[i + 1] = q_out0; dest[i + 2] = (int16_t)(acc1 >> 15); dest[i + 3] = q_out1; } cnv->old_x = ox; cnv->old_y = oy; cnv->old_e = oe; cnv->fir_head = f_head; cnv->dly_head = d_head; }